Memory

Every conversation with the AI Companion starts fresh, except for memory. Memory is a small set of durable facts about you and your work that the companion carries from one conversation to the next, so you do not have to re-explain your role, your territory, or how you like things done. Memory is per user, per team. The same person on a different team has a separate memory, so nothing about one team’s work leaks into another.

​

What a memory looks like

A memory has a type, a short name, and a one-sentence body written in the third person. The name describes what the memory covers, not the fact itself: “Role and team”, not “User is an enterprise SDR”. That is what lets a later change update the memory in place instead of creating a duplicate beside it.

​

How memories get saved

​

Automatically

After each turn, a background extractor reads the exchange and decides whether it revealed anything worth keeping. Most turns yield nothing; that is the expected outcome. A fact is saved only if all three hold:

1. It will still be true and useful next week, in a conversation about something else.
2. It is about you or your standing world, not about what you happen to be doing right now.
3. It is not already known from your profile, the platform, or the current conversation.

Some things are never saved automatically, because they are the most common false positives:

• Task requests. “Source ten SDR managers in San Francisco” is a job to do now, not a fact about you, however detailed it is.
• Rules inferred from one incident. One failed tool call or one odd result is not a standing preference.
• Transient state. An expired connection, a schedule that came back a certain way, something you are about to do.
• Platform knowledge. What an integration can or cannot do goes stale and is not about you.
• Anything already on your profile or visible in the conversation.

Before saving, the extractor compares the candidate against your existing memories. A fact that is already there, even worded differently, is skipped or merged, never duplicated.

​

Explicitly

Say it. “Remember that I only call the West Coast after 11am” saves a feedback memory immediately. “From now on…” and “always…” work the same way. You can also ask the companion to update a memory that is now wrong, to forget one, or to tell you what it remembers.

​

In the Memory tab

Open the companion and switch to Memory. Memories are grouped by type, and each one shows whether you saved it or the companion inferred it. You can add a memory with the form, edit any memory’s type, name, or body, and delete any memory.

​

How memories are used

At the start of every conversation, your memories are placed in the companion’s context under “About you”, and the companion treats them as things it already knows. When you have many memories, the ones not pre-loaded are still reachable: the companion searches memory when a question touches something it does not have in front of it.

​

Turning memory off

The Memory tab has a switch. When memory is off:

• Nothing new is saved, automatically or explicitly. The companion’s memory tools are removed from its toolset entirely.
• Existing memories are not shown to the companion and the tab reads “Memory is off”.
• Your memories are kept, not deleted, so turning it back on restores them.

​

Limits and housekeeping

When the cap is reached, the least valuable memory is retired to make room. Value is weighted mostly by importance, with explicit saves and standing rules ranked above one-off projects, and then by how recently and how often a memory has been used. A retired memory is archived rather than deleted immediately, and is recoverable for 90 days; saving a memory with the same name during that window brings it back.

​

Related